变形图驱动的人体与脸部网格模板拟合系统

为了建立统一的人体模型数据库/人脸模型数据库,实现了一个用标准三维人体/人脸网格模型拟合扫描数据的系统.该系统为人体模型构建特征点候选区域,用马尔可夫网络对候选区域进行概率预测得到人体特征点的位置;对于人脸模型,采用基于回归树的算法在多幅不同角度的人脸图像上分别检测其二维特征点,融合得到准确的三维特征点位置;得到特征点后,使用相似变换对标准模板和扫描模型的位置朝向以及尺度进行统一;通过特征点引导的变形图算法对标准模板和扫描模型的形状和姿态进行粗配准;最后使用基于稠密点对应的顶点仿射变换拟合得到变形后的标准模板.对CAESAR数据集中的人体模型以及扫描得到的人体和人脸模型均进行了拟合,实验结果表明,该系统能够精确地捕捉扫描数据的几何形状.

A 3D Human Body and Face Fitting System Based on Deformation Graph and Mesh Template

To build a unified human body model database / face model database, we propose an efficient system using a standard 3D human body mesh / face mesh model to fit the scanning data, so that the different shape models in the database have the same connectivity. It constructs a candidate area for detecting landmarks, and then uses a Markov network to detect the landmarks. As for human face, it employs a regression tree based algorithm detect 2D landmarks in face images captured from different viewing angles. These 2D landmarks are then fused to compute accurate 3D facial landmarks. After that, a similarity transformation is performed to align the standard template and the scanning model. Then the shape and pose of the standard template and the scanning model are roughly registered by using the deformable graph algorithm guided under feature points. Finally, the deformed standard template is fitted by vertex affine transformation based on dense point correspondence. The system is used to fit the human body models in the CAESAR dataset as well as the raw human body and face models. Experimental results show that the system can capture the geometry of the given data accurately.